Latest posts Visit blog
In plain words: what is this page about?

Many makers send the facts about their goods only as a document. It lists size, weight and power needs. In the shop, these facts are then missing from the filters. A program with artificial intelligence can read the documents. It writes each value into the right field. It also notes where in the document each value was found. A person checks unclear values before they appear in the shop. We show how this works.

Anyone selling technical products in B2B trade knows the pattern: the manufacturer supplies item number, name and price as a list, while the actual attributes sit in the PDF datasheet. Rated voltage, degree of protection, dimensions and material are there, but not as fields a shop can filter, compare or write into a feed. AI datasheet extraction means taking these details from text and tables as attribute, value and unit, mapping them to your own attribute schema and storing them with their source. The underlying technology has arrived in German companies: 26% of companies in Germany with at least ten employees used AI in 2025, and 52% of those used text mining, on a par with generating images, video and sound (Destatis). This article shows how a datasheet becomes a reviewed record in the PIM and why uncertain values belong in an approval step instead of going straight into the shop. We handle the implementation as part of our data enrichment with AI.

Why technical attributes often exist only in a PDF

Structured catalog data is the ideal case in B2B trade. BMEcat and ETIM are established formats in which manufacturers deliver class, attribute, value and unit in machine-readable form; how this handover works is described in our article on BMEcat and ETIM in catalog exchange. In practice, however, not every supplier provides such a file. Smaller manufacturers, imported goods, older items and spare parts often arrive with only a price list and a datasheet. With a range drawn from many sources, the result is a patchwork: some items have clean attributes, others only a file attachment.

For the shop, this is more than a cosmetic flaw. A filter for rated voltage or degree of protection shows only the items where the attribute is maintained as a field; all others drop out of the results even though they would fit. Product comparison stays incomplete, internal search does not find the value, and product feeds lack the technical details buyers search for. In B2B e-commerce in particular, such values decide the purchase: does the voltage match, is the degree of protection sufficient for outdoor use, does the dimension fit the installation?

A technical datasheet is usually well structured, just for people rather than machines. Each of its building blocks has its own sources of error:

  • Header data: manufacturer, type designation and item numbers, often for a whole product series.
  • Technical data table: attribute, value and unit in rows, with one column per variant for product series.
  • Dimension drawing: dimensions as annotations in a graphic, not as text.
  • Standards and markings: test marks, conformity statements and references to standards.
  • Footnotes and conditions: values that apply only with restrictions, for example without accessories or at a specific ambient temperature.
  • Notes in running text: installation, safety and care instructions, often spread over several pages.
Datasheet and catalog file complement each other

Even where a supplier delivers a catalog file, the datasheet is worth a look: catalog files are not necessarily fully populated. Extraction can fill gaps in a targeted way instead of overwriting data that has already been checked.

What AI can and cannot do when extracting data

Text mining, the analysis of written language, is the technology that matters when extracting data from datasheets. Among AI users it is widespread: 52% of companies in Germany that used AI in 2025 used text mining, 51% of companies with 10 to 49 employees and 57% of companies with more than 250 employees (Destatis). AI use itself rises with size: 57% of companies with more than 250 employees used AI, compared with 23% of those with 10 to 49 employees (Destatis). The population is companies with at least ten employees; from the 2025 reference year, the Federal Statistical Office delimits them using the EU definition of an enterprise.

Across Europe the share is lower: 19.95% of enterprises in the EU with at least ten employees used AI in 2025 (Eurostat). There was no predominant technology; text mining was used slightly more often than the others, by 11.75% of all enterprises (Eurostat). Among large enterprises with 250 or more employees, the share for text mining was 35.04%, higher than for any other AI technology recorded (Eurostat). The statistics calculate differently: Destatis relates the text mining share to companies that use AI, Eurostat to all enterprises.

When extracting, a language model recognizes pairs of attribute name and value in text and tables, separates number and unit and assigns both to a field in the target schema. This also works with changing layouts, with details in running text (“The housing is made of die-cast aluminum”) and with tables whose columns each manufacturer labels differently. What the model does not do: it does not know whether a value is correct. It can misread, overlook a footnote or fill a gap with a plausible value. That is why every value needs the location where it appears, and a check before it shows up in the shop. The following overview compares the three usual approaches.

CriterionManual typingFixed extraction rulesAI extraction with approval
New supplier with its own layoutpossible right away, but slownew rule needed per layoutusually no new rule needed
Details in running textare readhardly capturedare recognized
Tables with variant columnserror-prone with many columnsonly with a fixed structurepossible, check assignment to item number
Units and notationsdepends on the personaccording to fixed rulesnormalized according to fixed rules
Traceabilityrarely documentedrule is documentedsource stored per value
Effort at large volumesgrows with every itemlow as long as the layout stayslow, approval only for uncertain values
No source, no value

A language model produces an answer even when the document contains none. For product data this is the most dangerous type of error, because an invented value looks exactly like a correct one. The rule is therefore: a value without a verifiable location in the document is not accepted but reported as missing.

The target schema first: which attributes the shop needs

Before a datasheet is extracted, the question is what should end up in the record. The target schema defines for each product group which attributes exist, which data type they have, in which unit they are stored, which values are permitted and which serve as filters. If you classify according to ETIM, you can largely adopt this structure: ETIM describes a feature by description, feature type, unit and value and distinguishes four feature types, namely alphanumeric with a fixed value list, logical, numeric and range (ETIM International). Numeric and range features there need a unit of measurement, with the exception of plain counts such as “number of …” (ETIM International). Without ETIM, the schema is created in the PIM or in the shop’s properties, as described in our guide to PIM implementation.

  • Alphanumeric with value list: material “aluminum”, color “anthracite”. The datasheet often says “die-cast alu”; the schema decides which list value is meant.
  • Logical (yes or no): “dimmable”, “with power plug”. If the datasheet does not state it, the value is unknown, not “no”.
  • Numeric: rated voltage 230 V, weight 1.2 kg. Number and unit are stored separately.
  • Range: operating temperature from −20 to +50 °C. Lower and upper limit are two values with the same unit.

Most of the work lies in mapping the terms. Manufacturers name the same attribute differently, and similar terms do not necessarily mean the same thing. One datasheet says “operating voltage 220–240 V”, another “rated voltage 230 V”. Whether both belong in the same field is decided by the schema, not by the language model: a voltage range is a different attribute from a single value. A maintained mapping table per product group has proven its worth; it maps manufacturer terms to schema attributes and grows with every approval.

What is not in the schema is not invented

Extraction fills only attributes the schema provides for. Further details from the datasheet go into a list of notes for product management. This way the schema grows deliberately, and the record stays free of fields nobody maintains.

From PDF to record: attribute, value, unit, source

The process can be divided into eight steps. How to check in advance whether a PDF has a text layer and what happens to its tables is covered in the article checking PDFs for text layer, tables and scans for a knowledge base; that check is worthwhile here too. What differs is the goal and the handling of scans: instead of text passages for answers, the result is a structured record, field by field, and a scan is not rejected but read via text recognition and checked more strictly.

  1. Assign the document: link the datasheet to the supplier, the product series and the item numbers, and record the edition and date.
  2. Read text and tables: extract the text layer, keep tables as rows and columns, read scans via text recognition.
  3. Form candidates: recognize pairs of attribute name and value, including in running text.
  4. Map to the schema: assign the manufacturer term to a schema attribute via the mapping table.
  5. Normalize the unit: separate number and unit, standardize the notation, set the unit code.
  6. Store the source: record file, page, table, row and the raw text.
  7. Check the rules: verify value range, unit, contradictions and mandatory attributes.
  8. Accept or submit: accept reliable values, submit uncertain ones for approval, request missing ones from the supplier.

A common source of error lies in the first step. Many datasheets describe a product series: a table with one column per variant, with the type designations above it. Extraction has to assign each value to the right item number, and the bridge there is usually the type designation. Without a clean mapping from type to item, the 30-watt variant gets the value of the 20-watt variant. The same care applies to packaging and order units, which may be named differently in the datasheet than in the shop; how these sit in the item master is shown in the article on pack sizes and minimum order quantities.

The result of a run is one candidate per attribute, with its origin. The following example shows which fields such a candidate should carry; names and values are freely chosen.

candidate-weight.json
{
  "item": "LX20-AL-830",
  "attribute": "weight",
  "value": 1.2,
  "unit": "kg",
  "unit_code": "KGM",
  "raw_text": "Gewicht 1,2 kg*",
  "source": {
    "file": "datenblatt-lx20-de.pdf",
    "page": 2,
    "table": "Technische Daten",
    "row": 4
  },
  "note": "footnote: without mounting bracket",
  "confidence": "uncertain",
  "status": "approval_pending"
}

Two fields deserve special attention. The raw text records what literally appeared in the document, including asterisks and notation; the reviewer can see immediately whether normalization has changed anything. And the note carries conditions the value alone does not show: a weight “without mounting bracket” is a different value from the shipping weight, and a performance figure “at 25 °C” does not apply to every installation site.

Normalizing units without distorting values

Units are not a formatting detail. In Germany, quantities in official and commercial dealings must be stated in legal units where units for them have been set by statutory ordinance, and the specified names and unit symbols must be used for these units (EinhZeitG). For extraction, this sets a clear direction: what is stored is “1.2 kg” with the symbol kg, not “1.2 kilo” or “1.2 KG”. Conversions belong in a fixed rule with documented rounding, for example when a datasheet states grams and the schema expects kilograms.

For passing data on, a code helps more than free notation. Schema.org provides the UN/CEFACT Common Code with three characters for units (schema.org), for example KGM for kilogram, CMT for centimeter or VLT for volt. The record holds the code, and the shop displays the familiar symbol. This prevents the same unit from appearing in the filter in three notations. The following table shows typical raw values and their normalized form.

Raw value in the datasheetStored asWhat to watch for
850 g0.85 kg (KGM)conversion with documented rounding
230V~230 V ACcurrent type as a separate component, not as a symbol
24 VDC24 V DCkeep direct and alternating voltage apart
IP 65IP65standardize the notation, do not change the value
-20/+50 °C−20 to 50 °C (range)lower and upper limit as two values
Ø 120 mmdiameter 120 mmtranslate the symbol into the right attribute
Why the current type is not a detail

The Low Voltage Directive applies to electrical equipment designed for use with a voltage rating of between 50 and 1000 V for alternating current and between 75 and 1500 V for direct current, except for the equipment listed in Annex II (Directive 2014/35/EU). A device with 60 V falls within this range for alternating current but below it for direct current. If the current type is missing from the record, the value is not just imprecise, it leaves a legally relevant question open. Such attributes therefore belong in the approval step whenever the datasheet does not state the current type unambiguously.

Uncertain values belong in an approval step

Not every extracted value is equally reliable. A value from a clearly labeled table row with a unit is something different from a number taken from a poorly legible scan or a detail the datasheet states differently in two places. That is why each candidate receives a status: reliable, uncertain or missing. Reliable values go into the record after the rule checks, uncertain ones into an approval list, missing ones into a request to the supplier. Typical rules that detect uncertain values are these:

  • Value range: is the value within the permitted range for the product group? A light fitting with 2300 W is more likely a reading error than a product.
  • Unit matches the quantity: a weight in millimeters or a voltage without a unit stands out.
  • Contradiction in the document: table and running text state different values.
  • Condition in a footnote: the value applies only with restrictions, for example without accessories.
  • Deviation within the series: a value differs sharply from sibling items of the same series.
  • Mandatory attribute empty: an attribute required by the product group is missing from the document.

The approval itself should be quick, and the source makes sure of that: next to the proposed value, the reviewer sees the excerpt from the datasheet and confirms with one click or corrects it. Every correction flows back, into the mapping table or into the rules. The roles follow what has also proven itself in AI-assisted order entry: the machine proposes, fixed rules check, and a person decides where rules are not enough.

Google states a similar principle for the product feed: an attribute name and value should only be provided when the value is confirmed (Google Merchant Center Help). A default value such as “no” for a logical attribute is not a confirmation. The same applies to the shop, because a wrong filter value leads to mistaken purchases and returns, while a missing one only means one item fewer in the results.

Where the attributes take effect: filters, comparison, feed

In the shop, the benefit shows in three places. Filters need normalized values so that “230 V” and “230V~” do not appear as two entries and numeric attributes can be filtered as a range. Product comparison needs the same attributes in the same unit across all items of a product group. And search finds items by attribute values that were previously only in the attachment. How filter pages affect indexing is covered in the article on faceted navigation and filter URLs.

For Google Shopping, the “product detail” attribute (product_detail) is intended for technical specifications. It consists of three sub-attributes, namely section name, attribute name and attribute value (Google Merchant Center Help). Each detail may have 1 to 150 characters, and the field can be repeated up to 100 times (Google Merchant Center Help). How the feed can also be extended with questions, answers and links to manuals is shown in the article on conversational attributes in Merchant Center; we handle the connection through our Google Merchant Center integration.

Filters and facets

Normalized values and units produce clean filters, and numeric attributes can be filtered as a range.

Product comparison

The same attributes in the same unit make items from different manufacturers comparable.

Feeds and catalogs

Attributes flow into product feeds, to marketplaces and as a catalog file to business customers.

Mandatory information: responsibility stays with the retailer

The number of manufacturer documents arriving as files rather than on paper is likely to keep growing. Under the new Machinery Regulation, the instructions for use may be provided in digital form (Regulation (EU) 2023/1230); the regulation applies from 20 January 2027 (Regulation (EU) 2023/1230). For retailers of technical products, the question of how manufacturer information from PDFs gets into the shop in structured form therefore gains weight.

For products intended for consumers or likely to be used by consumers, the General Product Safety Regulation is added. It has applied since 13 December 2024 (Regulation (EU) 2023/988) and requires the online offer to include, among other things, information allowing the identification of the product and any warnings or safety information (Regulation (EU) 2023/988). Such notes are often found in the datasheet. Extraction can find them and assign them to the item; whether a note is complete and suitable is decided by a person. What Article 19 requires in detail is summarized in our article on GPSR in online shops.

What the statistics show about the hurdles

Introducing AI for product data rarely fails because of the technology alone. Of the companies in Germany that considered using AI in 2025 but did not implement it, 72% cited a lack of knowledge as a reason, and 44% cited difficulties with the availability or quality of the necessary data (Destatis). This group makes up 20% of all companies with at least ten employees (Destatis). Across Europe, a lack of relevant expertise was the most common reason, at 70.89%, among enterprises that had considered AI (Eurostat).

With datasheets, both can be defused. The necessary knowledge lies less in the language model than in the target schema and the rules, and your own product management knows both from daily work. Data quality is also a result here: extraction makes visible which attributes are missing and which suppliers deliver incomplete documents.

Introduction in five steps

  1. Choose a product group: start with a group where filters matter and many suppliers deliver only datasheets.
  2. Define the target schema: determine attributes, data types, units, value lists and mandatory fields, along the ETIM classes if needed.
  3. Run a pilot: extract datasheets from a few suppliers and compare the result with a manually checked sample.
  4. Set up rules and approval: define the rules, provide an approval view with the source, clarify responsibilities.
  5. Expand and maintain: connect further product groups, detect new datasheet versions by the stored edition and keep extending the mapping table.

A few key figures show whether the effort pays off: the share of items with all mandatory attributes filled, the share of values with a source, the rate of approvals without correction and the time from receipt of a datasheet to the approved record. When the rate of approvals without correction rises, schema and mapping table are ready for the next product group.

How we implement datasheet extraction

We build extraction as part of the product data pipeline, not as an isolated solution. It starts with the target schema and mapping table for the first product group. We then set up extraction with sources, the rules and an approval view, in the existing PIM or directly in the shop. The handover runs through our PIM system integrations or directly into the Shopware shop. Each solution is developed individually for your range and your suppliers, not as an off-the-shelf package.

Schema and mapping

Attributes, units and value lists per product group, plus a mapping table for manufacturer terms.

Extraction with source

Attribute, value and unit from text and tables, stored with file, page and raw text.

Approval and handover

Rules, an approval view for uncertain values and the handover to PIM, shop and feed.

From datasheet to filterable attribute

Datasheets will remain an important source in B2B trade for the foreseeable future, even as more manufacturers deliver structured catalogs. Extracting them with AI yields attributes for filters, comparison and feeds without typing every line. The order is what matters: first the schema, then extraction, and a source for every value. What the same care looks like for mandatory information is shown in the article on textile labeling online; a broader look at texts, translations and attributes is given in our overview of product data enrichment.

Sources and studies

This article is based on the tables of the Federal Statistical Office on the use of AI and on reasons against it (Destatis, reference year 2025, as of 24 November 2025), the Eurostat article on the use of AI in enterprises (extracted December 2025), the model description by ETIM International, schema.org (PropertyValue, QuantitativeValue, unitCode), the Google Merchant Center Help on the product_detail attribute, Directive 2014/35/EU, Section 1 EinhZeitG and Regulations (EU) 2023/1230 and (EU) 2023/988. Retrieved on 3 November 2026. The values in the graphic, tables and code are examples.

For legible datasheets with clear tables, usually yes; with scans, footnotes and variant tables, reliability drops. What matters is the process: every value is stored with its source, fixed rules check it, and uncertain values go into an approval step before they appear in the shop.

Not necessarily. ETIM provides a ready-made schema for technical products and makes catalog exchange easier, but a schema of your own in the shop works as well. To start, it is enough to define the target schema for one product group. More on this in the article on BMEcat and ETIM.

Scans are read via text recognition. With clean originals this usually works well; with low resolution or stamps, the error rate rises. Values from scans therefore get a stricter check, and asking the supplier for a digital version is often the faster route.

Number and unit are stored separately and converted, with documented rounding, into the schema’s unit. For passing data on, a unit code according to UN/CEFACT, which schema.org also provides for, is suitable. The raw text is kept so that every conversion can be traced.

The retailer is usually responsible for the information in its own offer, even if the data comes from the manufacturer. That is why legally relevant attributes such as voltage, degree of protection or safety notes belong in the approval step. This does not replace legal advice in an individual case.

That depends on the number of attributes, the quality of the datasheets and whether a schema already exists. A reliable estimate is possible after a short pilot with real datasheets. Get in touch with us via our contact page.